On 06/24/14 07:42, Alan Lawrence wrote:
I'm looking at git commit ea1ac559 / svn r76965, January 2014 (archive:
https://gcc.gnu.org/ml/gcc-patches/2004-01/msg03406.html), which prevents
(ashiftrt (xor A C1) C2)
from being commuted to
(xor (ashiftrt A C2) (ashiftrt C1 C2))
and wondering if anyone can explain to me what's wrong with this
transformation. Having worked through all four cases of A and C1
positive and negative, it seems to me that the extra bits 'fed in' to
the most-significant end of the result are the same either way (i.e. the
XOR of the sign bits of A and C1).
Re-enabling this transformation improves rtx simplification and codegen on
AArch64, for one.
(I don't have easy access to Alpha/VMS machine on which to test Ada
s-crc32.adb.)
You might be able to set up a cross compiler to see if/how this changes
the generated code.
Probably best to start with a compiler from that era since lots of stuff
may have changed and effectively hidden the issue Kenner was trying to
deal with.
I haven't heard from Kenner in a while, but you could always try to
contact him directly and see if he can recall the issue. It was a long
time ago though...
jeff
